FlashVLA · LingBot-VLA · RoboTwin 2.0

Streaming Action Decoding for Fast and Asynchronous VLA Inference

A LingBot-VLA flow-matching vision-language-action policy finetuned on RoboTwin 2.0 (50-task multitask) and served with FlashVLA streaming action decoding for fast, asynchronous inference.

Base model: robbyant/lingbot-vla-4b-posttrain-robotwin
Method: FlashVLA, streaming action decoding for flow-matching VLAs
Benchmark: RoboTwin 2.0, 50-task multitask (clean / randomized)

Results

RoboTwin 2.0 50-task multitask results. Success rate is averaged across clean and randomized settings. d=0 is synchronous execution; d=1 launches inference one step before the current chunk ends. Inference latency is independent of d and is therefore reported only at d=0.

Model	Avg SR (%)	Time/Step (ms)	Inference Latency (ms)
LingBot-VLA (base)	85.2	56.7	70.6
+FlashVLA (`d=0`)	88.6 (↑3.4)	49.5 (1.15× faster)	25.1 (2.81× faster)
+FlashVLA (`d=1`)	89.3 (↑4.1)	46.6 (1.22× faster)	—

Usage

Install FlashVLA:

git clone https://github.com/z-lab/flashvla.git
cd flashvla
conda env create -f environment.yml
conda activate flashvla

RoboTwin 2.0 evaluation uses a server/client split across two environments. See sim_eval/robotwin/ for setup and full options.

License

These weights are finetuned from robbyant/lingbot-vla-4b-posttrain-robotwin, which does not currently declare a license. Until terms are settled with the base model authors, no license is asserted for these weights and they should be treated as available for research and evaluation only. Please contact us before any other use.

The FlashVLA inference and training code is separately released under the Apache 2.0 License.

Citation

@inproceedings{li2026flashvla,
  title     = {{FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference}},
  author    = {Li, Zekai and Tang, Jiaming and Liu, Zhijian},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}